Crowdom is a tool for simplifying data labeling.
Write plain Python code and launch data labeling without knowledge of crowdsourcing and underlying platform (Crowdom uses Toloka as a platform for publishing tasks for workers). Define task you solve and load source data with few lines of code, choose quality-cost-speed tradeoff in interactive UI form, launch data labeling, study result labeling in Pandas dataframes.
We recommend you to look first at image classification example, since it demonstrates full data labeling workflow, proposed in Crowdom, with detailed explanations for each step.
In other examples, you can see how working with data labeling looks like for different types of tasks with use of Crowdom.
Join our Telegram chat if you want to learn more about the Crowdom or discuss your task with us.
Tasks in Crowdom are divided into two types:
- Classification tasks, which have a fixed set of labels as output.
- Annotation tasks, for which output has "unlimited" dimension.
In a typical classification task, worker is proposed to make a choice of one of the pre-determined options. Side-by-side (SbS) comparison is a special case of classification task.
As for annotation task, there may be many potential solutions, and there may be more than one correct one. Speech transcription, image annotation are examples of annotation tasks.
The following table contains list of examples, which demonstrates data labeling for different types of tasks, as well as other aspects of data labeling workflow.
Examples are presented as .ipynb files, located in this repository, but displayed by nbviewer,
which do it more precisely than GitHub.
Image classification and audio transcript) examples also have .html versions. These examples present full
labeling workflow, corresponding two classification and annotation types of tasks respectively. .html allows
to collapse optional sections in notebook to simplify understanding of main steps of workflow, as well as to display
interactive widgets contents (for example, to display quality-cost-speed tradeoff interactive form).
| Example | Full workflow | Function | Data types | Additionally |
|---|---|---|---|---|
| Image classification (HTML) | ✅ | Classification | Image | |
| Audio transcript (HTML) | ✅ | Annotation | Audio, Text | |
| Audio transcripts SbS | SbS | Audio, Text | ||
| Voice recording | Annotation | Text, Audio | Media output, checking annotations by the ML model | |
| Audio transcript, extended | Annotation | Text, Audio | Custom task UI, custom task duration calculation, first annotations attempts by the ML model | |
| MOS | Classification | Audio | MOS algorithm usage example | |
| Audio questions | Classification | Audio | Output label set depending on the input data | |
| Experts registration | Registration of your private expert workforce | |||
| Task update | Task update (instructions, UI and etc.) | |||
| Async launch | Non-blocking or parallel labeling launches |
Join our communities if you have questions about Crowdom or want to discuss your data labeling task.
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent